A Novel Optimization Procedure in WRBNN for Time Series Forecasting

The forecasting procedure based on wavelet radial basis neural network is proposed in this paper. The MODWT result becomes multivariate input of model. The smooth part constructs main pattern of forecasting model. Meanwhile the detail parts constuct the fluctuation rhythm of disturbances. The model assumes that the main pattern of model can be approximated by linear terms, meanwhile the fluctuation rhythm is nonlinear and will be approximated by nonlinear (radial basis) function. The LM test is used for exploring the number of wavelet coefficient clusters in every transformation level, which refer to the number of significant (optimum) radial basis node. The membership of cluster is decided by k-means method. The least square method is used for model parameters estimation.